Reduce code duplication in audio collection + some small fixes#15587

Open

racoiaws wants to merge 13 commits intoNVIDIA-NeMo:mainfrom

racoiaws:audio-refactor

Commits on Apr 7, 2026

Simplify SchroedingerBridge _step to return scalar loss

racoiaws
and
claude
committed
Extract _parse_batch helper into AudioToAudioModel base class

racoiaws
and
claude
committed
Move training_step into AudioToAudioModel base class

racoiaws
and
claude
committed
Guard SB component loss logging with self.training check

racoiaws
and
claude
committed
Move sample_rate and setup_optimization_flags to base AudioToAudioModel.__init__

racoiaws
and
claude
committed
Remove redundant world_size init from EncMaskDec and BNR2

racoiaws
and
claude
committed
Use self.from_config_dict in EncMaskDecAudioToAudioModel

racoiaws
and
claude
committed
Update setup_optimization_flags docstring

racoiaws
and
claude
committed
Extract _normalize/_denormalize helpers into base class

racoiaws
and
claude
committed
Remove misleading -> tuple annotation from _normalize

racoiaws
and
claude
committed
Fix CodeQL warning: use pass instead of ... in abstract method

racoiaws
and
claude
committed
Fix SB test that calls _step outside Lightning training loop

racoiaws
and
claude
committed
Fix _denormalize to be proper inverse of _normalize

racoiaws
and
claude
committed